The Prompt

Hello reader,

Last week, OpenAI admitted that its models had hacked Hugging Face during an internal cybersecurity test. Hungging Face had previously disclosed the attack without knowing who was responsible, using AI to detect the intrusion and analyze more than 17,000 recorded actions.

Both companies ended up working together, as they showcased how quickly AI is changing the cybersecurity game. AI can search for weak points and carry out attacks faster than humans, but it also helps defenders understand and respond to attacks at the same speed.

Those capabilities are becoming easier to access. Last week, Google released Gemini 3.5 Flash Cyber, a lightweight model trained to find, validate, and patch vulnerabilities. On the same day, Cisco introduced Antares, a series of small open-weight models that companies can repeatedly run over their own codebases to locate known flaws.

These releases point to a future where companies can evaluate their software's security more frequently and at a much lower cost. Being open-weight, Cisco's models can run within a company's own systems, allowing sensitive code to stay in-house rather than being sent to a third-party provider.

This matters for Southeast Asia, a long-standing hub for business process outsourcing and software services. Companies in these sectors must meet the security standards of international clients, and some of them will have to do so while relying on old or heavily customized systems that are difficult to update without disrupting operations.

As AI makes vulnerability discovery more routine, the bottleneck shifts to fixing those flaws. That creates opportunities for startups that don't just produce more security warnings, but help companies address them.

Elsewhere in this issue, we've updated our tracker of Chinese tech founders in Southeast Asia, adding 21 new names and a list of investors backing their companies. The tracker now covers 63 companies, mostly based in Singapore.

We also speak with Chrissy Lim, co-founder and CEO of Singapore-based Protaigé, about which parts of marketing AI can replace and which ones still need humans.

Glenn Kaonang, journalist

TOP AI READ FROM OUR DESK

Uncovering the elusive Chinese tech founders dominating SEA

Chinese influence in Southeast Asian tech has moved beyond the likes of Alibaba, Tencent, or ByteDance expanding into the region. Several Chinese entrepreneurs are building some of Southeast Asia's biggest startups and raising capital from both Chinese and global investors.

Beyond where these companies are headquartered, this list shows how deeply these founder networks shape the region's tech ecosystem. And with most of these founders building AI startups, they play a large role in dictating Southeast Asia's position in the global AI race.

WHAT WE THINK

1️⃣ Will Washington's accusation actually stop anyone from using Kimi?

The White House has accused Moonshot AI of using outputs from Anthropic's Claude Fable 5 to develop Kimi K3. It's not the first time a Chinese AI lab has faced allegations of "large-scale distillation" before, but it's the first time the accusation came from the top.

While the US hasn't made its evidence public, it's reportedly considering sanctions. But even if Moonshot is penalized, would that stop companies from using Kimi K3? Probably not entirely.

Kimi K3's full weights are expected to become downloadable today. Once that happens, the model becomes separable from the company that created it. Developers can run it themselves or use an outside company to host it.

Imagine a Singapore startup using Kimi K3 to serve American customers. It doesn't necessarily have to send customer information to Moonshot in China. It could access the model through an American inference provider such as Fireworks.

In fact, this setup has created a monetization tension among some Chinese AI labs, as we previously wrote.

What this means is even if the sanction required US inference providers to stop supporting the model, there would still be plenty of providers hosting the model elsewhere. The US can make it harder for Moonshot to do business, but it will struggle to disrupt the open-weight revolution.

2️⃣ As coding gets cheaper, startups may spend more on distribution

Chinese models such as DeepSeek and GLM have already made capable coding assistance available at very low prices. A new model called Laguna S 2.1 by US-based Poolside brings the price down further while focusing specifically on software development.

Laguna S 2.1 costs US$0.10 per million input tokens and US$0.20 per million output tokens, even cheaper than DeepSeek V4 Flash. It's also open-weight and small enough to run on one Nvidia DGX Spark - still a very expensive desktop computer at US$4,699 apiece, but one that can be easily run at home.

Pricing matters because coding agents can consume a lot of tokens while they read files, test code, spot errors, and redo tasks. Lower costs allow one founder to keep an agent working through longer tasks without worrying that every attempt is burning through their limited budget.

That could be particularly useful in Southeast Asia, where early-stage capital has become harder to raise. While we don't have strong evidence of solo founders booming in the region yet, cheaper tools could make bootstrapping more practical and allow founders to get further before raising money.

This may also change where startup budgets go. Cheaper software development means more people can launch competing products. Some of the budget that would've been used for engineering may go to marketing and customer onboarding efforts instead.

FOUNDER FOCUS

We speak with AI founders and executives about AI, startups, and whatever else is on their minds. In this edition, Tech in Asia sits down with Chrissy Lim, co-founder and CEO of Singapore-based Protaigé, an autonomous AI marketing platform.

Its AI account director, Maia, can take a campaign from brief to strategy and multichannel production. But Lim says the aim isn't to replace the creative spark - it's to remove the "80% perspiration" that follows.

The interview has been edited for brevity and clarity.

What was breaking in the traditional agency model before generative AI came along?

The economics have always been upside down. The hard part is cracking the brief: finding the insight, coming up with the big idea, and understanding the customer. But perhaps 20% of a project fee goes toward that thinking. Most of it gets absorbed by production, where the costs and headcount are.

At the same time, customers are spread across more channels and touchpoints, while attention spans are shrinking. Agencies are also far removed from the customer. There is the creative agency, the media agency, the marketer, the performance market, and client servicing. By the time a signal gets back to the creative team, it may be two or three weeks later, with little time or budget left to respond.

We all want to be agile, but traditional processes make that very difficult. The business of creativity is what kills creativity.

What happens when a marketer briefs Maia?

Everything begins with context. When a client comes on board, Protaigé learns about the brand, the market it operates in, and the customer it serves. We call that Brand DNA.

Then the marketer gives Maia a brief: what they need to launch, sell, or fix, and the outcome they want. Maia combines the brand's internal context with external signals, such as what is happening in the market or what competitors are doing. She turns that into a strategic and communications brief, then orchestrates the production of assets for different channels, markets, and customer segments.

Nothing is published without the user approving it. Human oversight matters most at the start of a campaign, when the core story and creative direction are being set. Once the strategic, conceptual, and design guardrails are in place, Maia can be trusted with more of the adaptation work.

Why build an end-to-end platform instead of a tool that specializes in copy, images, or videos?

It's like choosing between opening a chicken rice stall and a food court. One satisfies a craving, the other fulfills a broader need.

Marketing is a coherent, continuous conversation with a customer. Point solutions create another layer of fragmentation: someone has to take a copy from one tool, paste it into a design tool, reload the brand context, and then move everything into a publishing tool. The insights from one campaign need to feed back into the next so the system gets better over time.

That doesn't mean Protaigé has to build every underlying model. A specialist such as Runway may be better for producing a two-minute brand film scene by scene. Our role is to take that hero asset, adapt it, and distribute it across channels and markets.

We don't replace the 20% that is inspiration - we replace the 80% that is perspiration.

What parts of creative work still need humans then?

Coming up with the idea is still really hard. There is instinct and intuition involved in knowing whether an idea is novel, relevant, or simply not good enough. It is taste, empathy, and feeling.

AI can generate a million options, but a person still has to know which one is right. Technology is based on patterns from the past. Humans can invent something completely new.

We shouldn't expect AI to get everything right all the time. Even in an agency, we never get it right on the first try. It's still a collaborative process, except now you are working with a partner that is very fast, very bright, and improves the more you use it.

How does adopting Protaigé change a marketing team's day-to-day work?

Marketers spend a lot of time reviewing creatives, socializing them internally, giving feedback, and briefing and rebriefing. When the system already has the brand context, those review cycles can compress from weeks to minutes.

The time and budget saved can go toward getting closer to customers: doing more market research, holding events, creating more face-to-face contact, or testing new ways to reach people.

We have onboarded about 1,000 brands globally, including B2B software companies and agencies. Some agencies use Protaigé as a mini production studio, while others set up a campaign and let clients create adaptations themselves.

I built Protaigé for the agency I wanted to run: one where people get to do the fun part - the big idea, the first design, the first film - without the soul-destroying thousands of changes that follow.

What does AI ultimately replace in the agency business?

It's not about who the technology replaces but which part of the process it replaces. A digital campaign can require 50 or 60 banner sizes. Traditionally, someone has to adjust every one whenever there is a change. A system can now do that almost instantly.

That doesn't have to eliminate creative jobs. When computers entered studios, manual typesetting disappeared, but people became typographers and font designers. Generative AI creates a similar opportunity for agencies to build new experiences and learn new skills.

What will change is how the work is priced. Agency retainers were built around billable hours, and then the industry shifted toward projects and deliverables. If AI can continuously produce and optimize work toward a result, clients may pay for outcomes instead.

Retainers can survive. The question is: What are they retaining?

Catch up on past issues of the newsletter here. In the meantime, if you have any feedback or ideas, feel free to get in touch with Huong, our managing editor, at huong@techinasia.com.